The Law Society has called for clear safeguards as artificial intelligence develops from a tool that produces content into one that can make decisions and act without constant human instruction.
The justice system in England and Wales is not ready for the arrival of agentic artificial intelligence, the Law Society has warned, arguing that technology capable of taking multi-step action without repeated human instruction could deepen rather than relieve existing institutional weaknesses. Its new report, The future of agentic AI in the justice system, says autonomous systems may help reduce delay in routine administrative work, including document handling, triage, scheduling and enforcing decisions after a case has concluded. Yet it cautions that underfunding, fragmented data and still paper-heavy processes leave courts ill-equipped to deploy them safely.
The intervention lands at a time when Crown Court cases are being listed as far ahead as 2030, while civil courts are already grappling with litigants relying on inaccurate AI-produced legal material. The report says the central issue is not whether technology can produce faster results, but which parts of a justice system must remain the responsibility of a human being.
In a statement accompanying the report, Law Society vice president Brett Dixon said: “The justice system must not sacrifice fairness, accountability and public trust in pursuit of unmonitored efficiency.”
He continued: “AI is not a silver bullet to improve the justice system. It could help to ease some administrative pressures, but it is not a replacement for much-needed investment in the court estate and additional court staff.”
Beyond ChatGPT
Most people’s experience of AI will be generative tools such as ChatGPT, which produce text, summaries or other content in response to a prompt. The user asks a question, checks the answer and decides what to do next.
Agentic AI is an entirely different beast. It can be given an objective and then plan, sequence and carry out a chain of actions on behalf of a user or institution. In a legal setting, that could mean reviewing case papers, deciding the next procedural step, preparing documents, negotiating or progressing a claim, and triggering the next stage without a lawyer or court official prompting it each time.
The report describes this seismic shift as a move from AI as a tool used by people to AI acting as an agent for people or institutions. That distinction becomes more acute where a system’s decisions can affect an individual’s rights, access to a hearing or route through the court process.
There is no evidence, the Law Society says, of fully autonomous agentic AI currently being used in any justice system. Existing systems in the UK, US, Estonia and China are largely assistive or rules-based, with humans retaining formal authority. China’s Internet Courts come closest to a more autonomous model, although judges remain formally responsible for decisions.
Early UK examples nevertheless show the direction of travel. The report refers to a platform which uses agentic capabilities to manage small claims by automating workflow, claim progression, negotiation and document generation, with a qualified solicitor checking outputs. It says the adequacy of such supervision will become an increasingly live issue for firms using the technology.
Under pressure
The potential gains from automation are undeniable. Agentic AI could take on high-volume, low-discretion tasks and allow judges, lawyers and court staff to concentrate on contested facts, legal argument and the exercise of discretion. Properly deployed, it could make procedural help cheaper and easier to obtain for people unable to afford a lawyer – although, even when the less powerful agentic AI steps in, it is not without its pitfalls. But the report questions whether helping more people to start claims is enough. Access to justice, it says, also requires a fair hearing, a lawful decision and an effective remedy. A faster route into a system which cannot reliably deliver those things would not close the justice gap.
The danger is compounded by scale. A single fabricated authority, wrongly identified document or erroneous factual assumption may be caught when one lawyer uses a generative AI tool for a discrete task. In a chain of agentic systems, however, that error could be treated as reliable information by the next system and then passed onward through the process. The report therefore warns that even an apparently modest error rate could cause widespread harm when applied across thousands of cases. Corrective work and appeals could also consume any savings made at the outset.
It is particularly concerned that human oversight may become little more than ceremonial box-ticking. A judge or lawyer asked to approve an output from a complex, multi-step process may be unable in practice to understand and challenge what the system has done. The report calls this the risk of a ‘rubber-stamp’ model: nominal human sign-off without genuine human control.
What cannot be delegated
The Society does not contend that every use of agentic AI should be prohibited. Rather, it urges clear boundaries, transparency over how systems reach or influence outcomes, meaningful human oversight and public debate before autonomous tools become embedded in legal processes.
Interviewees involved in the research were unanimous that certain judicial functions should not be handed to an agentic system: deciding disputed facts, assessing witness credibility, exercising sentencing discretion, determining family cases and setting legal precedent. These are not simply tasks which current AI performs imperfectly and which will improve over time, the report says. They are functions whose legitimacy depends on a human decision-maker taking personal responsibility. Even a factually correct conclusion may lack legitimacy if the person affected cannot identify where a human being listened to their case and made the judgment.
Dixon said: “Legal judgement cannot be delegated. Not in courts and not in the legal sector. Professional expertise that shapes, checks, questions, improves and approves AI outputs is needed now more than ever.”
He added: “The most important question becomes not what AI can do, but what should remain human. There is still time to shape the boundaries of AI before the market does it for us.”
There is no immediate prospect of human judges abrogating their responsibility to decide cases in favour of robot judges. But there is a sense of the danger of mission creep where, case by case and contract by contract, automation becomes the ordinary route through the justice system – cheap, quick and increasingly difficult to challenge.
That may work tolerably well for routine administration. It is a far more troubling prospect when someone is facing eviction, loss of contact with a child, a criminal sentence or a dispute they cannot afford to take further. The Law Society is warning that, unless the boundaries of what agentic AI can and cannot do are decided now, the market may set them instead – by which point it may already be too late.